China has put generative artificial intelligence at the core of its military planning, and the Pentagon says Beijing is catching up with Washington. A Chinese large language model named Tianji has helped the People’s Liberation Army shrink a key combat decision loop to eight seconds.
The tool was built by DataExa, a company headquartered in Xiamen, and entered military service in 2023. It supports planning, force allocation, intelligence analysis and operational optimization. A comprehensive analysis was published on July 8 on the Substack newsletter “China: Technosphere.”
Tianji is offered in several sizes, from 7 billion to more than 130 billion parameters. The largest version probably requires six Huawei Ascend 910 processors with 64 GB of high-bandwidth memory each, and the architecture allows simultaneous deployment of multiple LLM modules and integration with DeepSeek.
How the PLA’s AI command tool works
DataExa, founded in 2014 by Hong Wanfu, is located in the Software Park II district in Xiamen’s Siming area. Its customers include the PLA, military universities and research institutes. The system optimizes logistics, plans missions and analyzes weapons data for the PLA command.
The model is trained on open-source intelligence, signals intelligence and human intelligence, as well as historical conflicts, military strength indexes and enemy tactics. According to the July analysis, the PLA’s Joint Operations Command Center used Tianji to shorten the chain of threat assessment, target assignment and battle damage assessment to eight seconds.
DataExa’s design resembles American military AI work: Palantir for structuring defense data, OpenAI’s Codex for coordinating multiple AI agents and Anduril’s Lattice for preparing tactical data. Tianji’s AI agents also create advanced wargaming simulation environments. Outside the military, DataExa serves the People’s Bank of China and China Mobile, and in March 2026 it released MilClaw, a military version of the OpenClaw autonomous AI agent.
Chinese labs are distilling American AI
China’s military ecosystem is also using US models to build its own systems. On July 31, Reuters reported on Jamestown Foundation research covering more than 80 Chinese papers and patents. The study found that scientists linked to the PLA and other security bodies used data from OpenAI and Anthropic models for reconnaissance, cyberwarfare, drones and tactical operations.
The key method is model distillation: outputs from large AI systems are used to train smaller models that can operate locally. In one documented example, a PLA intelligence and cyberwarfare unit used OpenAI’s GPT-3.5 to summarize sensitive military source code, then trained a national model to run exclusively on Chinese military networks. At North University of China in Taiyuan, a school tied to the country’s arms industry, researchers used Anthropic’s Claude 3 Haiku to generate synthetic training data for a social media monitoring and content moderation model.
Other projects use Western AI for live image analysis, drone navigation during communication failures and target recognition in simulated naval operations with drones and unmanned submarines. Because Washington restricts Chinese access to advanced chips, Beijing is pushing “model slimming” and edge computing so AI can run on drones, satellites and other devices with limited computing power.
Pentagon and think tanks take notice
In December 2025, the US Department of Defense report on Chinese military progress acknowledged that China is catching up in generative AI for military applications. The report listed unmanned systems, intelligence collection, decision support, cyber operations and information campaigns, and warned about AI-generated fake content for influence operations.
In May, the South China Morning Post revealed the Air Target Agent System, which combines large language models with AI agents for automated satellite targeting and reconnaissance. The system can decompose complex tasks, select algorithms, coordinate workflows and recover from failures on its own, moving from image recognition to conclusions and autonomous actions.
Georgetown University’s Center for Security and Emerging Technology reviewed PLA procurement requests and identified similar priorities: decision-support systems, sensor upgrades and algorithms that generate, expand and merge data to speed up decisions and improve military precision.
China is also looking at how to defend against AI reverse engineering. In January, scientists from the Army Engineering University published research on “data-free distillation,” a technique for recreating a model’s abilities without access to its internal parameters, and proposed ways to hide logical information exposed in public outputs.
The military dimension of US-China AI competition is becoming more visible. Openly available models are of limited use to armed forces because of the risk of leaking classified data, but understanding their inner workings allows the creation of similar tools designed for military purposes. The ability to influence how AI systems think may eventually matter as much as understanding human decision-makers.
Source: WNP.PL, Photo: DC Studio / Shutterstock
